U.S. patent application number 15/326511 was filed with the patent office on 2017-07-20 for compact multifunctional system for imaging spectroscopy.
This patent application is currently assigned to Ecole Polytechnique Federale de Lausanne (EPFL). The applicant listed for this patent is ECOLE POLYTECHNIQUE FEDERAL DE LAUSANNE (EPFL). Invention is credited to Yosef Akhtman, Dragos Constantin.
Application Number | 20170205337 15/326511 |
Document ID | / |
Family ID | 54151327 |
Filed Date | 2017-07-20 |
United States Patent
Application |
20170205337 |
Kind Code |
A1 |
Akhtman; Yosef ; et
al. |
July 20, 2017 |
Compact Multifunctional System for Imaging Spectroscopy
Abstract
A method for obtaining spectral imaging data comprises at least
the steps of receiving a sample set of data generated by sampling a
spectral property of an image of an object in a spatial basis,
wherein the sampling of the spectral property of the image of the
object comprises providing a Spectral Filter Array (SFA) by
arranging a plurality of SFA elements together to form a surface;
configuring each SFA element of the plurality of SFA elements to
filter one or more spectral bandwidths centered each at specific
wavelengths corresponding to that SFA element, whereby all of the
plurality of SFA elements taken together cover a determined
spectral range; and setting the specific wavelengths of each SFA
element of the plurality of SFA elements on the surface such to
obtain a uniform and aperiodic spatial distribution of all of the
plurality of SFA elements across the surface. The sampling of the
spectral property of the image of the object further comprises
providing an image sensor configured to record at each pixel the
light filtered by one of the plurality of SFA elements or a subset
of the plurality of SFA elements thereby producing one intensity
value of light filtered by the one of the plurality of elements or
the subset of the plurality of SFA elements per pixel; forming the
image of the object on the SFA through a lens or group of lenses;
and recording for all of the pixels of the image sensor the
spectrally filtered intensity values thereby obtaining a
2-dimensional array of the intensity values corresponding to the
image of the object. The method for obtaining spectral imaging data
further comprises the step of reconstructing a full 3 dimensional
spectral data cube of the imaged object from the sampled
2-dimensional array.
Inventors: |
Akhtman; Yosef; (Saint-Prex,
CH) ; Constantin; Dragos; (Lausanne, CH) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
ECOLE POLYTECHNIQUE FEDERAL DE LAUSANNE (EPFL) |
Lausanne |
|
CH |
|
|
Assignee: |
Ecole Polytechnique Federale de
Lausanne (EPFL)
Lausanne
CH
|
Family ID: |
54151327 |
Appl. No.: |
15/326511 |
Filed: |
July 24, 2015 |
PCT Filed: |
July 24, 2015 |
PCT NO: |
PCT/IB2015/055614 |
371 Date: |
January 16, 2017 |
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G01J 3/0229 20130101;
G06K 2009/00644 20130101; G01N 21/31 20130101; G01J 3/027 20130101;
G01J 2003/2826 20130101; G06K 9/0063 20130101; G01J 3/2823
20130101; G06T 17/05 20130101; G06K 9/00657 20130101; H04N 13/214
20180501 |
International
Class: |
G01N 21/31 20060101
G01N021/31; G06T 17/05 20060101 G06T017/05; H04N 13/02 20060101
H04N013/02; G06K 9/00 20060101 G06K009/00 |
Foreign Application Data
Date |
Code |
Application Number |
Jul 24, 2014 |
IB |
PCT/IB2014/063366 |
Claims
1. A method for obtaining spectral imaging data, comprising the
steps of: receiving a sample set of data generated by sampling a
spectral property of an image of an object in a spatial basis,
wherein the sampling of the spectral property of the image of the
object comprises providing a Spectral Filter Array (SFA) by
arranging a plurality of SFA elements together to form a surface,
configuring each SFA element of the plurality of SFA elements to
filter one or more spectral bandwidths centered each at specific
wavelengths corresponding to that SFA element, wherein all of the
plurality of SFA elements taken together cover a determined
spectral range, and setting the specific wavelengths of each SFA
element of the plurality of SFA elements on the surface such to
obtain a uniform and aperiodic spatial distribution of all of the
plurality of SFA elements across the surface; providing an image
sensor configured to record at each pixel the light filtered by one
of the plurality of SFA elements or a subset of the plurality of
SFA elements thereby producing one intensity value of light
filtered by the one of the plurality of elements or the subset of
the plurality of SFA elements per pixel, forming the image of the
object on the SFA through a lens or group of lenses, recording for
all of the pixels of the image sensor the spectrally filtered
intensity values thereby obtaining a 2-dimensional array of the
intensity values corresponding to the image of the object,
reconstructing a full 3 dimensional spectral data cube of the
imaged object from the sampled 2-dimensional array.
2. The method of claim 1, wherein the step of reconstructing
comprises a mathematical modelling and computational numerical
optimization.
3. The method of claim 2, wherein the mathematical modelling and
computational numerical optimization comprises a machine learning
method which enables an absence of transmission measurement by the
SFA.
4. The method of claim 2, wherein the mathematical modelling and
computational numerical optimization comprises at least a convex
optimization method based on transmission measurement of the
SFA.
5. The method of claim 1, wherein the step of reconstructing
comprises providing measured transmissions of individual SFA
elements of the SFA, creating a system design matrix from measured
transmissions of individual SFA elements and image sensor
radiometric calibration, and inferring non sampled spectral
information using deconvolution or non-linear sparse reconstruction
methods based on the system design matrix.
6. The method of claim 1, wherein the step of setting the specific
wavelengths of each SFA element of the plurality of SFA elements on
the surface such to obtain the uniform and aperiodic spatial
distribution of the specific wavelengths of the SFA is
deterministic, thereby comprising an aperiodic tiling such as Wang
or Penrose.
7. The method of claim 1, wherein the step of setting the specific
wavelengths of each SFA element of the plurality of SFA elements on
the surface such to obtain the uniform and aperiodic spatial
distribution of the specific wavelengths of the SFA further
comprises filling the SFA surface with a repeating pattern, then
continuously and randomly inter-changing the SFA elements until the
measured entropy of the SFA central wavelengths is sufficiently
high.
8. The method of claim 1, wherein the step of setting the specific
wavelengths of each SFA element of the plurality of SFA elements on
the surface such to obtain the uniform and aperiodic spatial
distribution of the specific wavelengths of the SFA further
comprises a random sampling of the SFA elements from a uniform
distribution.
9. The method of claim 1, wherein the step of setting the specific
wavelengths of each SFA element of the plurality of SFA elements on
the surface such to obtain the multiband distribution of the
specific wavelengths of the SFA further comprises a sampling of
multiband SFA elements such that neighboring SFA elements are
spectrally orthogonal.
10. A method for Spectral Filter Array (SFA) element second-degree
transmission cancellation through SFA design and subsequent
response subtraction comprising at least: designing a uniformly and
aperiodically distributed Spectral Filter Array (SFA) by arranging
a plurality of SFA elements together to form a surface, configuring
each SFA element of the plurality of SFA elements to filter a
spectral bandwidth centered at a central wavelength corresponding
to that SFA element, whereby all of the plurality of elements taken
together cover a determined spectral range, and setting the central
wavelength of each SFA element of the plurality of SFA elements on
the surface such to obtain a distribution of all the central
wavelengths to be uniform and aperiodic over the surface;
identifying the SFA elements with second degree transmission in the
sensitivity range of an image sensor, changing the central
wavelength of SFA elements neighboring the SFA elements with second
degree transmission, as to match the central wavelength of the
second degree transmission, re-arranging the SFA element central
wavelengths which were not changed or do not have a second degree
transmission, as to maintain the uniformity of the SFA central
wavelength distribution, and in a 2-dimensional array of pixel
intensity values, subtracting the response corresponding to the
changed neighboring elements from the response corresponding to the
SFA elements which have a second degree transmission.
11. A method for obtaining a real-time monochromatic preview of an
imaged object, the imaged object having been obtained using the
method of obtaining spectral imaging data of claim 1, the method
for obtaining the real-time monochromatic preview comprising the
steps of: designing a uniform aperiodic Spectral Filter Array (SFA)
by arranging a plurality of SFA elements together to form a
surface, configuring each SFA element of the plurality of SFA
elements to filter a spectral bandwidth centered at a central
wavelength corresponding to that SFA element, whereby all of the
plurality of SFA elements taken together cover a determined
spectral range, and setting the central wavelength of each SFA
element of the plurality of SFA elements on the surface such to
obtain a distribution of all the central wavelengths to be uniform
and aperiodic over the surface, generating a subset of SFA elements
by a 2-dimensional periodic selection, replacing the SFA elements
belonging to the above subset with SFA elements of identical
transmission, and from a 2-dimensional array of pixel intensity
values, using a subset of values, corresponding to the subset of
periodic SFA elements with identical transmission, to create a
lower resolution monochromatic image.
12. A method for obtaining a real-time monochromatic preview of an
imaged object, the imaged object having been obtained using the
method of obtaining spectral imaging data of claim 1, the method
for obtaining the real-time monochromatic preview comprising the
steps of: designing a uniform aperiodic Spectral Filter Array (SFA)
by arranging a plurality of SFA elements together to form a
surface, configuring each SFA element of the plurality of SFA
elements to filter multiple spectral bands centered at specific
wavelengths corresponding to that SFA element, whereby all of the
plurality of SFA elements taken together cover a determined
spectral range, and setting the central wavelength of each SFA
element of the plurality of SFA elements on the surface such that
neighboring SFA elements are spectrally orthogonal, from a
2-dimensional array of pixel intensity values, spatially
interpolating each channel corresponding to an SFA element
configuration independently at pixel locations where other SFA
element configurations are present, and averaging the independently
interpolated channels, to create a high resolution monochromatic
image.
13. A method for generating a spectral 3D-model of an imaged object
and increasing the spectral reconstruction quality of an imaged
scene, comprising: taking multiple images of the scene, each of the
images comprising at least an array of pixel intensities, using
monochromatic previews to optically align the images through image
registration, thus generating the 3D-model of the scene, grouping
values from different ones of the arrays of pixel intensities
corresponding to different images of the multiple images, if the
pixels are aligned to a same point in the imaged scene, based on
grouped values, re-modeling a system design matrix corresponding to
multiple arrays of pixel intensities, taken from different
locations, and reconstructing the spectral texture of the 3D model
based on the cross-image system design matrix.
14. A non-transitory computer readable medium having a
computer-executable program stored thereon, the computer-executable
program, when executed by a computing device, configured to perform
a method according to claim 1 to reconstruct a 3D-hyperspectral
image from a 2D-spatial-spectral dataset.
Description
TECHNICAL FIELD OF INVENTION
[0001] The invention relates to the field of imaging spectroscopy,
and more precisely hyperspectral imaging. More specifically, the
invention comprises a compressive spectral image acquisition and
reconstruction system.
BACKGROUND OF INVENTION
[0002] Imaging Spectroscopy
[0003] Imaging spectroscopy is a digital sensing process in which a
scene is optically sampled in two spatial dimensions and in one
spectral dimension, producing a three dimensional data cube.
Spectral images contain significantly larger amount of spectral
information than color photography, typically comprising tens or
hundreds of well defined narrow spectral channels for each
individual pixel of the image. In other words each spatial pixel of
the spectral image contains a spectral response of the respective
point of the imaged surface. In comparison, a color image is
comprised by three loosely defined Red-Green-Blue channels.
Spectral images further distinguish themselves from color
photography through the rigorously measured radiometric output:
while color is perception based and has no absolute units, spectral
radiance is a physical measure whose unit is [Wm.sup.-2
.mu.m.sup.-1 sr.sup.-1]. As a consequence, spectral imaging is
often used to determine the chemical and biological composition of
objects, while avoiding the need for physical contact [1, 2]. The
acquisition of 3D hyperspectral imaging data is difficult because
of the two-dimensional nature of the imaging sensors.
[0004] In order to obtain the spectral data cube with a two
dimensional sensor, most spectral imaging cameras use the following
three elements: [0005] 1. An optical element, such as a lens: to
focus the optical scene onto an imaging plane. [0006] 2. A
dispersive element, such as a prism or diffraction grid: to
spatially distribute spectral information from the imaging plane.
[0007] 3. An imaging sensor, such as a CCD or a CMOS: to spatially
sample the dispersed light.
[0008] Due to the spectral dispersion, spatial resolution must be
compromised. To overcome the resolution loss, many spectral imaging
cameras employ some form of scanning: [0009] Pushbroom, or
line-scan, cameras capture a spatio-spectral slice of the datacube,
using the two dimensions of the imaging sensor as 1D spectral and
1D spatial sampling. Pushbroom sensors require movement to
reconstruct an entire hyperspectral image, they require spatial
scanning. [0010] Band sequential cameras capture a spectral slice
of the spectral data cube, using either rotating or tunable
spectral filters. They use the two dimensions of the imaging sensor
as 2D spatial sampling and scan spectrally by sequentially applying
different spectral filters in front of the imaging sensor, they
thus require spectral scanning. [0011] Interferometry-based
spectral cameras sequentially capture interferograms of the data
cube onto the image sensor. The interferograms are generated by
tuning an interferometer, a process similar to spectral
scanning.
[0012] Frame, or snapshot, spectral cameras acquire the entire data
cube without scanning, severely compromising resolution.
State-of-the-art snapshot spectral cameras are either based on 2D
diffraction grids or small filter banks, requiring a high
resolution sensor and producing very low resolution spectral data
cubes with respect to scanning spectral cameras [2].
[0013] Compressive Sensing
[0014] Classical signal processing dictates that in order to
sample, then reconstruct a signal without information loss, the
signal has to be sampled with at least twice the highest frequency
it contains. This minimal sampling frequency is also called Nyquist
frequency. If sampling is done under the Nyquist frequency, the
reconstructed signal will show information loss and artefacts such
as aliases. The large number of measurements required to
reconstruct a full hyperspectral data cube under the Nyquist
constraint is one of the main reasons why all non-compressive
snapshot spectral cameras have low resolution.
[0015] Recent research in signal processing and advances in
computational power have led to many improvements in signal
acquisition bandwidth reduction through compressive sensing. At the
core of compressive sensing lays the hypothesis that if the signal
that needs to be acquired is sparse in some mathematical basis, the
number of measurements required to reconstruct the signal are given
by its sparsity and can thus be significantly smaller than the
number of measurements dictated by the Nyquist frequency.
[0016] The effective reconstruction of the original signal using
the compressive sensing mathematical models imposes certain
requirements on the underlying signal sampling methodology, in
particular uniformity and aperiodicity.
[0017] In areas of applications such as magnetic resonance imaging
(MRI), compressive sensing has been successfully implemented in
order to reduce the amount of required measurements by a factor of
between 10 and 20 [4], or, with a corresponding compression ratio
of between 1/10 and 1/20. For color images, the method has been
suggested in [5], but the practical implementation has been limited
by the low information gain and high computational cost with
respect to classical demosaicing. In color photography, the
required compression ratios is 1/3 as the images have a small
quantity of spectral information.
[0018] Compressive sensing has also been applied to spectral
imaging in multiple scientific experiments, showing compression
ratios of between 1/4 and 1/16, [6][7]. These experiments, however,
failed to produce the results needed for a practical snapshot
spectral camera which requires compression ratios around 1/50 or
more. The color imaging methods described in [5] cannot be directly
applied to 2D spectral sampling-based imaging for a plurality of
reasons: [0019] Correct pixel-wise spectral reconstruction poses a
series of new challenges such as optimal spectral distribution of
filters, or higher order filter response cancellation. [0020] Color
filter arrays do not provide sufficient spectral separation [0021]
For hyperspectral image reconstruction, the algorithms need to
reconstruct a significantly larger set of data from less than 5% of
its size, as opposed to 33% subsampling rate of color imaging.
[0022] Interferometric Spectral Filter Array
[0023] Due to the very stringent requirements of spectral imaging,
a spectral filter array is much more difficult to produce than the
color arrays used in commercial color cameras. Color cameras
generally employ pigment-based filters whose spectral transmission
bandwidth is too wide and irregular for the radiometric
measurements of spectral imaging. The number of different filters
that can be produced in a pigment-based color filter array is also
limited by the number of pigments used. The cost of a color filter
array thus greatly increases with the increase in the number of
different filters.
[0024] Interferometric filters, such as Fabry-Perot filters, are
based on a different principle than pigment filters and enjoy a
number of properties useful for imaging spectroscopy: [0025] finely
tunable central wavelength: the central wavelength is given by the
distance between the reflective surfaces used to make light
interfere with itself. Varying this distance changes the central
wavelength of the filter without significantly altering the shape
of its spectral transmission curve; [0026] symmetry around central
wavelength: the filter transmission is locally symmetric with
respect to the central wavelength, due to the natural
interferometric process; [0027] finely tunable bandwidth: the
filter bandwidth is given by the reflectivity of the surfaces which
reflect the light; [0028] the production cost of a large number of
different wavelength interferometric filters is much lower than for
pigment-based filters.
[0029] While their properties make them an obvious choice for
imaging spectroscopy, until recently, interferometric filters were
not feasible to create at the size of a single pixel. Fabry-Perot
filters were thus used to produce spectral images by
band-sequential scanning, as large global filters, which cover the
entire sensor surface [8]. However, recent developments in
miniaturization and nanotechnology have led to the development of
interferometric spectral filter arrays with the individual filters
being no larger than a pixel [4].
[0030] One of the main disadvantages of interferometric filters is
their second and higher order transmissions. These higher order
transmissions of the filters let light pass not only at the desired
central wavelength, but also at twice that wavelength, three times
that wavelength and so on. Interferometric filters can also be
configured as multiband filters, due to the higher order
transmissions which include multiple specific wavelengths for which
they transmit light. In most cases, the higher order transmissions
are undesired effects and need to be nullified.
SUMMARY OF THE INVENTION
[0031] In a first aspect, the invention provides a method for
obtaining spectral imaging data. The method comprises at least the
steps of receiving a sample set of data generated by sampling a
spectral property of an image of an object in a spatial basis,
wherein the sampling of the spectral property of the image of the
object comprises providing a Spectral Filter Array (SFA) by
arranging a plurality of SFA elements together to form a surface;
configuring each SFA element of the plurality of SFA elements to
filter one or more spectral bandwidths centered each at specific
wavelengths corresponding to that SFA element, whereby all of the
plurality of SFA elements taken together cover a determined
spectral range; and setting the specific wavelengths of each SFA
element of the plurality of SFA elements on the surface such to
obtain a uniform and aperiodic spatial distribution of all of the
plurality of SFA elements across the surface. The sampling of the
spectral property of the image of the object further comprises
providing an image sensor configured to record at each pixel the
light filtered by one of the plurality of SFA elements or a subset
of the plurality of SFA elements thereby producing one intensity
value of light filtered by the one of the plurality of elements or
the subset of the plurality of SFA elements per pixel; forming the
image of the object on the SFA through a lens or group of lenses;
and recording for all of the pixels of the image sensor the
spectrally filtered intensity values thereby obtaining a
2-dimensional array of the intensity values corresponding to the
image of the object. The method for obtaining spectral imaging data
further comprises the step of reconstructing a full 3 dimensional
spectral data cube of the imaged object from the sampled
2-dimensional array.
[0032] In a preferred embodiment the step of reconstructing
comprises a mathematical modelling and computational numerical
optimization.
[0033] In a further preferred embodiment, the mathematical
modelling and computational numerical optimization comprises a
machine learning method which enables an absence of transmission
measurement by the SFA.
[0034] In a further preferred embodiment, the mathematical
modelling and computational numerical optimization comprises at
least a convex optimization method based on transmission
measurement of the SFA.
[0035] In a further preferred embodiment, the step of
reconstructing comprises providing measured transmissions of
individual SFA elements of the SFA; creating a system design matrix
from measured transmissions of individual SFA elements and image
sensor radiometric calibration; and inferring non sampled spectral
information using deconvolution or non-linear sparse reconstruction
methods based on the system design matrix.
[0036] In a further preferred embodiment, the step of setting the
specific wavelengths of each SFA element of the plurality of SFA
elements on the surface such to obtain the uniform and aperiodic
spatial distribution of the specific wavelengths of the SFA is
deterministic, thereby comprising an aperiodic tiling such as Wang
or Penrose.
[0037] In a further preferred embodiment, the step of setting the
specific wavelengths of each SFA element of the plurality of SFA
elements on the surface such to obtain the uniform and aperiodic
spatial distribution of the specific wavelengths of the SFA further
comprises filling the SFA surface with a repeating pattern, then
continuously and randomly inter-changing the SFA elements until the
measured entropy of the SFA central wavelengths is sufficiently
high.
[0038] In a further preferred embodiment, the step of setting the
specific wavelengths of each SFA element of the plurality of SFA
elements on the surface such to obtain the uniform and aperiodic
spatial distribution of the specific wavelengths of the SFA further
comprises a random sampling of the SFA elements from a uniform
distribution.
[0039] In a further preferred embodiment, the step of setting the
specific wavelengths of each SFA element of the plurality of SFA
elements on the surface such to obtain the multiband distribution
of the specific wavelengths of the SFA further comprises a sampling
of multiband SFA elements such that neighboring SFA elements are
spectrally orthogonal.
[0040] In a second aspect, the invention provides a method for
Spectral Filter Array (SFA) element second-degree transmission
cancellation through SFA design and subsequent response
subtraction. The method comprises at least designing a uniformly
and aperiodically distributed Spectral Filter Array (SFA) by
arranging a plurality of SFA elements together to form a surface;
configuring each SFA element of the plurality of SFA elements to
filter a spectral bandwidth centered at a central wavelength
corresponding to that SFA element, whereby all of the plurality of
elements taken together cover a determined spectral range; and
setting the central wavelength of each SFA element of the plurality
of SFA elements on the surface such to obtain a distribution of all
the central wavelengths to be uniform and aperiodic over the
surface. The method further comprises identifying the SFA elements
with second degree transmission in the sensitivity range of an
image sensor; changing the central wavelength of SFA elements
neighboring the SFA elements with second degree transmission, as to
match the central wavelength of the second degree transmission;
re-arranging the SFA element central wavelengths which were not
changed or do not have a second degree transmission, as to maintain
the uniformity of the SFA central wavelength distribution; and in a
2-dimensional array of pixel intensity values, subtracting the
response corresponding to the changed neighboring elements from the
response corresponding to the SFA elements which have a second
degree transmission.
[0041] In a third aspect, the invention provides a method for
obtaining a real-time monochromatic preview of an imaged object,
the imaged object having been obtained using the method of
obtaining spectral imaging data described herein above, the method
for obtaining the real-time monochromatic preview comprising
designing a uniform aperiodic Spectral Filter Array (SFA) by
arranging a plurality of SFA elements together to form a surface,
configuring each SFA element of the plurality of SFA elements to
filter a spectral bandwidth centered at a central wavelength
corresponding to that SFA element, whereby all of the plurality of
SFA elements taken together cover a determined spectral range, and
setting the central wavelength of each SFA element of the plurality
of SFA elements on the surface such to obtain a distribution of all
the central wavelengths to be uniform and aperiodic over the
surface; generating a subset of SFA elements by a 2-dimensional
periodic selection; replacing the SFA elements belonging to the
above subset with SFA elements of identical transmission; and from
a 2-dimensional array of pixel intensity values, using a subset of
values, corresponding to the subset of periodic SFA elements with
identical transmission, to create a lower resolution monochromatic
image.
[0042] In a fourth aspect, the invention provides a method for
obtaining a real-time monochromatic preview of an imaged object,
the imaged object having been obtained using the method of
obtaining spectral imaging data described herein above, the method
for obtaining the real-time monochromatic preview comprising
designing a uniform aperiodic Spectral Filter Array (SFA) by
arranging a plurality of SFA elements together to form a surface,
configuring each SFA element of the plurality of SFA elements to
filter multiple spectral bands centered at specific wavelengths
corresponding to that SFA element, whereby all of the plurality of
SFA elements taken together cover a determined spectral range, and
setting the central wavelength of each SFA element of the plurality
of SFA elements on the surface such that neighboring SFA elements
are spectrally orthogonal; from a 2-dimensional array of pixel
intensity values, spatially interpolating each channel
corresponding to an SFA element configuration independently at
pixel locations where other SFA element configurations are present;
and averaging the independently interpolated channels, to create a
high resolution monochromatic image.
[0043] In a fifth aspect, the invention provides a method for
generating a spectral 3D-model of an imaged object and increasing
the spectral reconstruction quality of an imaged scene, comprising
taking multiple images of the scene, each of the images comprising
at least an array of pixel intensities; using monochromatic
previews to optically align the images through image registration,
thus generating the 3D-model of the scene; grouping values from
different ones of the arrays of pixel intensities corresponding to
different images of the multiple images, if the pixels are aligned
to a same point in the imaged scene; based on grouped values,
re-modeling a system design matrix corresponding to multiple arrays
of pixel intensities, taken from different locations; and
reconstructing the spectral texture of the 3D model based on the
cross-image system design matrix.
[0044] In a sixth aspect, the invention provides a computer program
stored on a memory device, the computer program when executed by a
computing device reading the memory device enabling to reconstruct
a 3D-hyperspectral image from a 2D-spatial-spectral dataset, by
implementing the steps of the method for obtaining spectral imaging
data as described herein above.
[0045] A corresponding hyperspectral image reconstruction module
comprises a single- or multi-processor computing device, or a
computer network programmed to reconstruct a 3D hyperspectral image
from a 2D dataset. The underlying algorithmic framework of the
hyperspectral image reconstruction may include deconvolution,
non-linear sparse reconstruction, regressive machine learning
methods, projection of the acquired spectral samples onto a set of
a priori known spectral signatures, as well as any combination of
the herein three methods. Algorithmic methods are also provided for
variable spatio-spectral resolution reconstruction of the imaged
scene, based on a fixed configuration imager.
BRIEF DESCRIPTION OF THE FIGURES
[0046] The invention will be better understood in view of the
description of preferred embodiments and in reference to the
drawings, wherein:
[0047] FIG. 1A is a schematic overview of the core imaging design
and process;
[0048] FIG. 1B is a variation of the 2D spectral sampling imager
based on a spectral filter array adapter fitted between the main
imaging lens and the imaging sensor;
[0049] FIG. 1C is a variation of the 2D spectral sampling imager
where the reconstructor is integrated into the imager;
[0050] FIG. 2A shows a color preview of an aerial image scene, the
corresponding sensor readout and the color preview of the
reconstructed 100 band spectral data cube, as computed by a
simulation. Four points of interest are marked on the reconstructed
image;
[0051] FIG. 2B-C shows a spectral comparison between the original
image scene sampled at 100 bands (marked "Original") and the
reconstructed spectral data cube (marked "Reconstructed"), for each
of the four points marked in FIG. 2;
[0052] FIG. 3A shows an example of a uniformly distributed and
aperiodic 12.times.12 element spectral filter array pattern given
by central wavelength in nanometers. 61 possible filters were
considered with central wavelengths between 400 and 100 nm, at a
step of 10 nm;
[0053] FIG. 3B is a variation of FIG. 3A where elements having a
non-negligible second degree transmission are always placed in the
direct vicinity of elements filtering at twice their central
wavelength. These filter pairs are highlighted in gray;
[0054] FIG. 3C is a variation of FIG. 3A where the aperiodic
pattern is interleaved with a periodic pattern of non-filtering
elements or elements having a common spectral transmission. The
elements of the periodic pattern are highlighted in gray;
[0055] FIG. 4 depicts an image filtering configuration where the
spectral filter array is preceded by an anti-aliasing filter and
followed by a color filter array, along the optical path of the
light to the imaging sensor;
[0056] FIG. 5A-B contain example plots of the continuous and
discretized spectral transmission of a Fabry-Perot spectral array
element with a central wavelength of 700 nm and a full width at
half maximum of 100 nm; and
[0057] FIG. 6 shows example plots of transmissions of two
Fabry-Perot spectral array elements designed for multiband
filtering. Variations in both multiband wavelength number and
bandwidth are illustrated. A wideband filter designed with 6
specific wavelengths (above) and a narrowband filter designed with
5 multiband wavelengths (below) are plotted.
DETAILED DESCRIPTION OF THE INVENTION
[0058] In this section, we outline the advantages of the presented
solution with respect to prior art and subsequently focus on the
technical details pertaining to three major aspects of the
invention, namely: [0059] 1. the camera design and functionality;
[0060] 2. the spectral filter array design; and [0061] 3. the
reconstructor embodiments and preferred implementation.
[0062] The present invention improves upon prior art in multiple
ways. With respect to spatial or spectral scanning spectral imaging
systems, the present invention overcomes the need to scan, sampling
the imaged scene simultaneously while providing similar resolution.
Due to the extremely short optical path of the present invention, a
much smaller and lighter spectral image acquisition system can be
built than those requiring spectrally dispersive elements. The
reduced complexity of the optical and mechanical parts also
significantly lowers the production cost when compared to spectral
imaging systems with dispersive elements. With respect to
state-of-the-art snapshot spectral acquisition systems, the present
invention generates significantly higher resolution spectral data
cubes. The present invention greatly reduces the amount of storage
required to acquire a spectral data cube as it is reconstructed
from the relatively small 2D set of optical samples that does not
exceed the number of pixels of the imaging sensor, while the full
data cube contains as many elements as the number of pixels on the
sensor multiplied by the number of registered spectral bands. The
size ratio between the 2D set of samples and the reconstructed
hyperspectral data cube can vary depending on the configuration,
but will typically be between 1/10 and 1/200.
[0063] Camera
[0064] A core embodiment of the invention comprises an imaging
spectroscopy camera system depicted in FIG. 1A, employed for
producing a 2D set of spectral samples 8 and a computational method
which is utilized to reconstruct a complete 3D spectral data cube
from the aforementioned 2D set of optical samples. More
specifically, the camera comprises an optical lens 5, or a group of
lenses, used for producing an image of the object of interest 1 on
the image plane. The camera further comprises a spectral filter
array (SFA) 6 wherein the various spectral components of the
object's image are transmitted to the image sensor, and the image
sensor array 7 arranged to detect the light transmitted by the
individual SFA elements. The camera also contains an
analog-to-digital signal converter and a storage medium 9, on which
the resultant 2D datasets are stored.
[0065] The corresponding spectral data cube reconstruction module
3, or reconstructor, comprises a single- or multi-processor
computing device, or a computer network programmed to reconstruct a
3D spectral image from the 2D set of optical samples.
[0066] The underlying algorithmic framework of the spectral image
reconstruction may include deconvolution, non-linear sparse
reconstruction, projection of the acquired spectral samples onto a
set of a priori known spectral signatures, as well as any
combination of the herein described three methods.
[0067] The imaging process is as follows: [0068] The imaged scene
is focused by the lens onto the plane of the SFA. [0069] Each
element of the SFA lets only certain frequencies of light pass
through it, depending on its transmission curve. [0070] Light
transmitted by the SFA falls onto the imaging sensor which
generates, for each pixel, an electrical charge in a manner
proportional to the amount of light received by the pixel. [0071]
The electrical charges in the imaging sensor are discretized and
digitized by an analog-to-digital converter, producing the sensor
readout. [0072] The sensor readouts are then stored onto the
storage medium. [0073] The reconstructor will receive the sensor
readouts from the storage medium and will reconstruct the full 3D
spectral image.
[0074] In some embodiments, the camera of this invention is based
on an interchangeable lens camera as depicted in FIG. 1B. The
spectral filter array 6 and a refocusing lens 14 are built into an
adapter 13 which fits in between the lens 5 and the imaging sensor
7 of the interchangeable lens camera. The image from the lens is
focused on the spectral filter array, then refocused onto the
imaging sensor by the refocusing lens. In these embodiments, any
interchangeable lens color camera becomes the core of this
invention by using such an adapter 13.
[0075] Various embodiments of the main imaging lens 5 include a
fixed focal optical lens, or lens groups, while others feature a
variable focal and variable focus lens.
[0076] Some embodiments see the storage medium 9 integrated in the
camera while others have no storage medium, the images being sent
to a computer or a computer network directly after acquisition and
analog-to-digital conversion.
[0077] Spectral Filter Array
[0078] Most embodiments of the present invention include a
uniformly distributed aperiodic spectral filter array 6, which we
refer to as SFA, composed of individual spectral filters, referred
to as elements. An example of such an SFA embodiment is shown in
FIG. 3A where each element is represented by its central
wavelength. In other implementations, the SFA may also contain
non-filtering elements, while preserving an aperiodic distribution
of interferometric filters across the majority of its surface, as
shown in FIG. 3C. In the preferred implementation, the SFA is
composed of interferometric filters such as Fabry-Perot
filters.
[0079] The uniform aperiodic distribution may be obtained in
several ways: [0080] deterministically through aperiodic tiling
such as Wang or Penrose tiling. [0081] by filling the SFA surface
with a repeating pattern, then continuously and randomly
inter-changing the SFA elements until the measured entropy of the
SFA central wavelengths distribution is sufficiently high. [0082]
by random sampling the SFA elements from a uniform
distribution.
[0083] Some embodiments of the invention include SFA designs where
the second and higher degree transmissions of the interferometric
spectral filters are nullified by means of several methods: [0084]
A method for second degree cancellation consists of placing the SFA
6 over an existing color filter array 11 (such as a Bayer pattern
filter array), matching the individual SFA elements' transmission
so that only the first order response will be transmitted by the
color filter array, as shown in FIG. 4. [0085] Another method for
second order transmission cancellation is always placing a SFA
element with its transmission equal to that of the second degree of
one of its adjacent elements, if those filters have 2nd degree
transmission in the sensitivity range of the sensor. The design
depicted in FIG. 3B, allows for subtracting the second degree
transmission based on adjacent filters, prior to
reconstruction.
[0086] Some embodiments of the invention use interferometric
filters with multiband wavelength transmissions, rather than
central wavelengths, as the elements of the SFA. Examples of these
multiband filter transmissions are shown in FIG. 6. In these
embodiments, the spatial uniform aperiodic distribution of the
elements is chosen to maximize the spectral orthogonality of the
multiband filters. By spectral orthogonality between filters we
mean that they respond differently to the same wavelengths.
Maximizing spectral orthogonality translates to choosing filters
which do not correlate spectrally.
[0087] In various embodiments, an element at periodic positions in
the SFA has the same transmission across the entire spectral filter
array as shown in FIGS. 4 and 1A. This periodic element can be
either an interferometric spectral filter, a pigment-based color
filter, or a non-filtering element transmitting all wavelengths of
light. The pixels under these periodic elements will be used to
create preview images 15 of the imaged scene. These images can be
used as real-time previews of the imaged scene or directly in image
registration and image based 3D reconstruction.
[0088] Some embodiments of the invention include a camera where an
anti-aliasing filter 12 is placed before the spectral filter array,
as depicted in FIG. 4, for homogenizing the light reaching the
different filters. The anti-aliasing filter may be comprised by:
birefringent filters extending over more than 2.times.2 of the
sensor elements, a defocused objective lens with a matched aperture
stop, and/or a degraded imaging lens.
[0089] Spectral Cube Reconstruction
[0090] In various embodiments of the invention, the reconstructor 3
is either a single computer, a programmable single or
multi-processor 10 device or a computer network programmed to
reconstruct the hyperspectral image from an acquired
two-dimensional dataset 8 through non-linear sparse reconstruction.
The reconstructor can also be a virtual machine, accessible through
the internet and ran on cloud-based computational servers.
[0091] Various embodiments feature the reconstructor integrated in
the camera and producing the hyperspectral image immediately after
having acquired the two-dimensional set of optical samples. FIG. 1C
depicts such an integrated camera, composed of the imaging optics 5
which focus the light onto the SFA 6 that is deposited over the
imaging sensor 7. Data from the imaging sensor 8 is then saved on
the storage medium 9 and converted to spectral data by the
reconstructor 3, composed of multiple processing units 10.
[0092] The mathematical model describing the sampling and
discretization of the imaged scene by the hyperspectral imager of
this invention is described by the following linear system:
Ax=y, (1)
where A is the system design matrix containing the spectral
transmission of the individual elements of the SFA, y is the sensor
readout, and x is the hyperspectral data cube corresponding to the
spatio-spectral properties of the imaged scene. Vectors x and y are
thus serialized versions of the 3D hyperspectral and 2D sensor
datasets.
[0093] While the SFA elements sample the continuous incoming light
spectrum, the design matrix A contains discrete versions of the SFA
transmissions, whose spectral resolution will be equal to the
number of spectral bands of the reconstructed data cube. This
property allows for the reconstruction of a various number of
spectral bands, this number having only a lower bound, dictated by
the minimum number of bands required to properly represent the SFA
element transmission curves. In FIGS. 5A and 5B, the effects of the
aforementioned discretization and lower bound are depicted: while
30 and 10 bands produce an adequate representation of the SFA
element transmission, 3 bands are insufficient.
[0094] Under the hypothesis of local spectral homogeneity of the
imaged scene, the number of spectral samples in y can be
artificially increased by pixel grouping, resulting in a
non-diagonal system design matrix A. This hypothesis can be further
enforced by the use of an anti-aliasing filter placed in front of
the SFA. Each spatial pixel of the image sensor readout constitutes
a single spectral sample of the complete spectral response of the
corresponding point in the imaged scene. The reconstruction process
may include the grouping of multiple spectral samples per pixel,
which are derived from the spatial pixels in the close vicinity of
the processed pixel.
[0095] One method for pixel grouping may be a sliding 2D window,
which groups all spectral samples from pixels falling inside its
area into the central pixel. If the SFA contains a periodic
interleaved pattern, another method for pixel grouping may be
guided by the contrast gradients observed in the dense image
directly obtained from the pixels behind periodic elements. Using
this dense image allows for grouping of pixels which do not span
across edges or high contrast areas, thus reducing the possible
artefacts produced by pixel grouping.
[0096] The variable resolution discrete filter representation and
the pixel grouping methods described above work in unison to allow
for the adjustment of the reconstructed hyperspectral data cube's
spatial and spectral resolutions. Depending on the desired
hyperspectral data cube resolution desired, only A and y need to be
recomputed accordingly, while the hyperspectral imager
configuration remains fixed.
[0097] Reconstruction of the hyperspectral cube can be achieved by
applying regressive machine learning methods [12] such as neural
networks or random forests to the sensor readout or pixel-grouped
sensor readout. These methods take in an input signal or feature
(y) and output an estimation of the original signal (x). Through
training with a large number of examples, models such as neural
networks can learn to un-mix and correct spectral data from the
sensor readout.
[0098] Another method for reconstructing the hyperspectral data
cube x from the measurements y can be modelled as a convex
optimization problem:
argmin(.parallel.Ax-y.parallel..sub.2+c(x)) (2)
where the function c(x) is an always positive penalty term,
constraining the search space for the optimal x based on a priori
knowledge of the properties of x. The function c(x) will have low
values for instances of x corresponding to a priori knowledge,
while .parallel.Ax-y.parallel..sub.2 will have low values for
instances of x fitting the measured data y. The convex optimisation
(2) can thus be interpreted as finding the spectral data cube x
which simultaneously best fits the measured data y and the a priori
knowledge about its structure c(x).
[0099] Variations of the aforementioned reconstruction model
include changing the representation of x to a mathematical basis in
which x becomes sparse. For instance, basis such as Fourier, direct
cosine transform, wavelet or gradient are known to be sparse for
natural images [9]. In these cases, (2) may become:
argmin(.parallel.ASx-y.parallel..sub.2+.lamda..parallel.Sx.parallel..sub-
.1) (3)
where S is the matrix representation of the transformation from the
chosen sparse mathematical basis and c(x) is replaced with the
L.sub.1 norm which enforces sparsity on the representation of x,
while A controls the strength of penalization. For hyperspectral
data cubes, other transformations can be envisioned, such as
projecting the measured data onto a known set of spectral vectors,
obtained through principal component analysis, for instance.
[0100] The preferred implementation of the reconstruction algorithm
is based on a version of the fast iterative shrinkage-thresholding
algorithm (FISTA) introduced in [10], using the total variation
(TV) norm as the penalty term. The reconstructed data cube is thus
found by minimizing:
argmin(.parallel.Ax-y.parallel..sub.2+.lamda..parallel.x.parallel..sub.T-
V) (4)
where .parallel.x.parallel..sub.TV can either be the 3D TV norm of
the entire reconstructed hyperspectral data cube, or the sum of the
2D TV norms of the individual spectral bands of the reconstructed
hyperspectral data cube. Using the sum of 2D TV norms rather than
the 3D TV norm has the advantage of independently processing
spectral bands and allowing for parallelization of the most
computationally intensive steps of FISTA.
[0101] If the measurements y were produced by a system
configuration having an anti-aliasing filter in the optical path,
the quality and speed of the reconstruction produced by FISTA can
be improved by independently blurring the spectral bands of the
estimated reconstruction at each iteration of the algorithm, in an
amount proportional to the blurring induced by the anti-aliasing
filter. This process can be parallelized.
[0102] A simulation of the imaging and reconstruction process of
the invention has been used to validate the principles of the
present invention, the results of which are presented in FIGS. 2A
and 2B. The target compression ratio was 1/100, meaning the
hyperspectral data cube was reconstructed from 100 times fewer
measurements than its total number of elements. A SFA model based
on Wang tiling and Fabry-Perot filter transmissions was employed,
containing 100 filters of central wavelengths between 400 and 1000
nm, and 100 nm full width at half maximum (FWHM). The SFA element's
transmissions were discretized in 100 spectral bands. The SFA also
contained a periodic pattern of elements having the central
wavelength at 700 nm, as shown in FIGS. 3C and 5. Gradient-based
pixel grouping was applied, grouping together 4 pixels from a
vicinity of 5.times.5 pixels. FISTA was configured to minimize the
sum of 2D TV norms of independent bands in parallel and included
the blurring step at each iteration.
[0103] The obtained simulated sensor readout as well as a color
preview of the hyperspectral data cube are shown in FIG. 2A. Four
points of interest have been marked on the reconstructed image,
presenting various spectral properties as well as being located in
areas of various spatial detail across the imaged scene. The
spectral properties of the original spatio-spectral image scene and
the reconstructed hyperspectral data cube are compared in these
four points, the results shown in FIGS. 2B and 2C. The spectral
resolution is notable, where the reconstructed spectrum exhibits
detail at much finer resolution than the FWHM of the SFA elements
used to sample the spectrum. If a spectral camera with 100 nm wide
filters at half-maximum had directly sampled such a spectrum, sharp
spectral features such as peaks would have been smoothed out by
these filters through the sampling process. However, in our
reconstructions we clearly see sharp peaks and valleys spanning
less than 10 bands (or 60 nm), well below the FWHM of the SFA
elements used for the spectral sampling.
[0104] Commercial Applications
[0105] Imaging spectroscopy technology has numerous and proven
applications in research and commercial domains, including
agriculture, natural resource management, mineralogy, medicine and
manufacturing among others. Specifically, the disclosed invention
facilitates the development of a new class of compact, lightweight
and inexpensive spectral imaging sensors. The new sensors are
suitable for deployment using small unmanned aircraft systems, and
cater for many of the existing applications, while facilitating a
fundamentally new range of usage scenarios.
[0106] In particular, in the context of the agricultural and
related industries the presented market size estimate is discussed
in detail in [11].
[0107] Airborne spectral imaging constitutes the single most
effective method of large-scale monitoring and analysis of
vegetation with a proven capability in: [0108] early detection,
diagnosis and control of plant diseases; [0109] stress detection
and growth monitoring; [0110] detection and control of invasive
species.
[0111] Despite the many proven benefits, today's airborne spectral
imaging technology is too expensive and complex for effective
exploitation in the agricultural and related industries. In
particular, none of the existing solutions has gained any
significant traction in commercial farming applications thus
far.
[0112] The proposed invention has the potential to make systematic
spectral monitoring of vegetation accessible and highly profitable
by reducing the cost of data acquisition and processing by a factor
of 10, or more, while providing up to 10% increase in yield and
associated revenues for the customers.
[0113] The available US statistics suggests 100 million hectares of
prime croplands that may require systematic monitoring at a rate of
5 times a year or more. Approximately 10,000 spectral imaging
systems are required in order to conduct the necessary monitoring.
The resultant combined value of hardware of data processing
services may be estimated as US $540 million for the USA market.
Assuming the size of the global market to be five times the size of
the USA market and taking into account other fields of application
including environmental monitoring, forestry, control of invasive
species, etc., results in an estimate for the total global market
for airborne spectral vegetation monitoring of at least $2
billion.
[0114] Summarizing, the invention relates to the field of imaging
spectroscopy, which is a method for optical sensing of both spatial
and spectral image properties. Specifically, the invention
describes a system and methodology to obtain spectral images, also
known as spectral data cubes due to their three-dimensional nature
comprised by two spatial and one spectral dimensions.
[0115] The described system is structurally simple and is designed
to facilitate the manufacturing of a compact and lightweight
spectral imaging camera. The system embodiment includes a lens, a
2D spectral filter array, a 2D imaging sensor, as well as the
storage and the data processing mechanisms required to obtain a
spectral image. Preferably the invention comprises the specific
configuration of the spectral filter array comprised by spectral
filters positioned in front of the individual light-sensitive
elements of the image sensor. The invention further describes a
computational framework utilized for the reconstruction of a 3D
spectral data cube from the 2D dataset provided by the imaging
sensor, as well as a real-time generation of a preview image.
Algorithmic methods are also provided for variable spatio-spectral
resolution reconstruction of the imaged scene, based on a fixed
configuration imager.
REFERENCES
[0116] [1] Ellis, J., (2001) Searching for oil seeps and
oil-impacted soil with hyperspectral imagery, Earth Observation
Magazine. [0117] [2] Lacar, F. M., et al., (2001) Use of
hyperspectral imagery for mapping grape varieties in the Barossa
Valley, South Australia, Geoscience and remote sensing symposium
(IGARSS '01)--IEEE 2001 International, vol. 6 2875-2877 p. [0118]
[3] Tack, K., Lambrechts, A., Haspeslagh, L. (2011) Integrated
circuit for spectral imaging system, Patent No. WO/2011/064403
[0119] [4] Lustig et al, (2007) Sparse MRI: The Application of
Compressed Sensing for Rapid MR Imaging, MRM 58:1182-1195 [0120]
[5] Singh, T., Singh, M. (2011) Method and System for Compressive
Color Image Sampling and Reconstruction. Patent no. US20110142339
A1. [0121] [6] A. Wagadarikar, R. John, R. Willett, and David
Brady, (2008) Single disperser design for coded aperture snapshot
spectral imaging, Appl. Opt. 47, B44-B51. [0122] [7] Golbabaee, M.,
Vandergheynst, P., (2012) Hyperspectral image compressed sensing
via low-rank and joint-sparse matrix recovery, IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP), p.
2741-2744 [0123] [8] Heikki, S., Ville-Veikko, A., Altti A., Tapani
A., Christer H., Uula K., Jussi M., Jyrki O., (2009) Novel
miniaturized hyperspectral sensor for UAV and space applications.
Proc. SPIE 7474, Sensors, Systems, and Next-Generation Satellites
XIII, 74741M, doi:10.1117/12.830284. [0124] [9] Subhasis, S.,
(2000) Image compression--from DCT to wavelets: a review.
Crossroads 6, doi:10.1145/331624.331630 [0125] [10] Beck, A.,
Teboulle, M., (2009) Fast Gradient-Based Algorithms for Constrained
Total Variation Image Denoising and Deblurring Problems, IEEE
Transactions on Image Processing, vol. 18, no. 11, pp. 2419,2434,
November 2009 doi:10.1109/TIP.2009.2028250 [0126] [11] The Economic
Impact of Unmanned Aircraft Systems Integration in the United
States, AUVSI, 2013 [0127] [12] T. M. Mitchell, "Machine Learning",
1997, ISBN: 0070428077 9780070428072
* * * * *